2 * Lockless get_user_pages_fast for s390
4 * Copyright IBM Corp. 2010
5 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 #include <linux/sched.h>
9 #include <linux/hugetlb.h>
10 #include <linux/vmstat.h>
11 #include <linux/pagemap.h>
12 #include <linux/rwsem.h>
13 #include <asm/pgtable.h>
16 * The performance critical leaf functions are made noinline otherwise gcc
17 * inlines everything into a single function which results in too much
20 static inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
21 unsigned long end, int write, struct page **pages, int *nr)
23 struct page *head, *page;
27 mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
29 ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
33 /* Similar to the PMD case, NUMA hinting must take slow path */
34 if (pte_protnone(pte))
36 if ((pte_val(pte) & mask) != 0)
38 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
40 head = compound_head(page);
41 if (WARN_ON_ONCE(page_ref_count(head) < 0)
42 || !page_cache_get_speculative(head))
44 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
48 VM_BUG_ON_PAGE(compound_head(page) != head, page);
52 } while (ptep++, addr += PAGE_SIZE, addr != end);
57 static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
58 unsigned long end, int write, struct page **pages, int *nr)
60 struct page *head, *page;
64 mask = (write ? _SEGMENT_ENTRY_PROTECT : 0) | _SEGMENT_ENTRY_INVALID;
65 if ((pmd_val(pmd) & mask) != 0)
67 VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
71 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
73 VM_BUG_ON(compound_head(page) != head);
78 } while (addr += PAGE_SIZE, addr != end);
80 if (WARN_ON_ONCE(page_ref_count(head) < 0)
81 || !page_cache_add_speculative(head, refs)) {
86 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
97 static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
98 unsigned long end, int write, struct page **pages, int *nr)
103 pmdp = (pmd_t *) pudp;
104 if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
105 pmdp = (pmd_t *) pud_deref(pud);
106 pmdp += pmd_index(addr);
110 next = pmd_addr_end(addr, end);
113 if (unlikely(pmd_large(pmd))) {
115 * NUMA hinting faults need to be handled in the GUP
116 * slowpath for accounting purposes and so that they
117 * can be serialised against THP migration.
119 if (pmd_protnone(pmd))
121 if (!gup_huge_pmd(pmdp, pmd, addr, next,
124 } else if (!gup_pte_range(pmdp, pmd, addr, next,
127 } while (pmdp++, addr = next, addr != end);
132 static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
133 unsigned long end, int write, struct page **pages, int *nr)
135 struct page *head, *page;
139 mask = (write ? _REGION_ENTRY_PROTECT : 0) | _REGION_ENTRY_INVALID;
140 if ((pud_val(pud) & mask) != 0)
142 VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
145 head = pud_page(pud);
146 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
148 VM_BUG_ON_PAGE(compound_head(page) != head, page);
153 } while (addr += PAGE_SIZE, addr != end);
155 if (WARN_ON_ONCE(page_ref_count(head) < 0)
156 || !page_cache_add_speculative(head, refs)) {
161 if (unlikely(pud_val(pud) != pud_val(*pudp))) {
171 static inline int gup_pud_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
172 unsigned long end, int write, struct page **pages, int *nr)
177 pudp = (pud_t *) pgdp;
178 if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
179 pudp = (pud_t *) pgd_deref(pgd);
180 pudp += pud_index(addr);
184 next = pud_addr_end(addr, end);
187 if (unlikely(pud_large(pud))) {
188 if (!gup_huge_pud(pudp, pud, addr, next, write, pages,
191 } else if (!gup_pmd_range(pudp, pud, addr, next, write, pages,
194 } while (pudp++, addr = next, addr != end);
200 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
201 * back to the regular GUP.
203 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
206 struct mm_struct *mm = current->mm;
207 unsigned long addr, len, end;
208 unsigned long next, flags;
214 len = (unsigned long) nr_pages << PAGE_SHIFT;
216 if ((end <= start) || (end > TASK_SIZE))
219 * local_irq_save() doesn't prevent pagetable teardown, but does
220 * prevent the pagetables from being freed on s390.
222 * So long as we atomically load page table pointers versus teardown,
223 * we can follow the address down to the the page and take a ref on it.
225 local_irq_save(flags);
226 pgdp = pgd_offset(mm, addr);
230 next = pgd_addr_end(addr, end);
233 if (!gup_pud_range(pgdp, pgd, addr, next, write, pages, &nr))
235 } while (pgdp++, addr = next, addr != end);
236 local_irq_restore(flags);
242 * get_user_pages_fast() - pin user pages in memory
243 * @start: starting user address
244 * @nr_pages: number of pages from start to pin
245 * @write: whether pages will be written to
246 * @pages: array that receives pointers to the pages pinned.
247 * Should be at least nr_pages long.
249 * Attempt to pin user pages in memory without taking mm->mmap_sem.
250 * If not successful, it will fall back to taking the lock and
251 * calling get_user_pages().
253 * Returns number of pages pinned. This may be fewer than the number
254 * requested. If nr_pages is 0 or negative, returns 0. If no pages
255 * were pinned, returns -errno.
257 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
265 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
266 * because get_user_pages() may need to cause an early COW in
267 * order to avoid confusing the normal COW routines. So only
268 * targets that are already writable are safe to do by just
269 * looking at the page tables.
271 nr = __get_user_pages_fast(start, nr_pages, 1, pages);
275 /* Try to get the remaining pages with get_user_pages */
276 start += nr << PAGE_SHIFT;
278 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
279 write ? FOLL_WRITE : 0);
280 /* Have to be a bit careful with return values */
282 ret = (ret < 0) ? nr : ret + nr;